In practical terms it matters very much. If the end point is
distributed,
multihoming it is a much more complex operation (assuming you want
sessions
to survive a multihoming event). Perhaps if you speak to some people
actually implementing virtualized server pools this would be clearer
to you.
This depends where we put the demarcation point. I think a place that
puts multihoming functionality on one side, and host virtuality on the
other would be a very good choice. In other words: this is something
for the cluster software implementers to worry about.